Blog

Securing the OT Data Path for AI

Securing the OT Data Path for AI

Key takeaways

  • Industrial AI is moving from analytics dashboards into production workflows, so the connection that feeds it data is now part of how the plant runs.

  • Most AI data paths use the same tools as remote access, including inbound firewall rules, VPNs, vendor gateways, and shared credentials.

  • Scope every AI connection to one process and one data source and keep the path that sends telemetry out separate from anything that changes operations.

  • Outbound-only, process-level connections let plants feed AI platforms without making the OT network reachable from the internet.

Most industrial AI projects start with a data problem. Teams spend weeks figuring out which historian holds the vibration readings, whether timestamps from three sites line up, and who has time to clean up a decade of tag names. Those questions deserve that attention. However, moving the data from the plant to wherever the model runs gets far less thought, and many teams solve it in an afternoon with a firewall exception and a service account.

That shortcut was easy to live with when AI meant a dashboard. AI now works inside production workflows, where it schedules maintenance, generates work orders, and recommends setpoint changes. The connection that feeds the model has become part of how the plant runs, and it deserves the same scrutiny as any other path into the control network.

Industrial AI is moving from advisor to operator

AI has run in OT for years, mostly as anomaly detection on a single machine or as a predictive model sitting next to the equipment it watched. Those systems stayed local, so their connections stayed small. Today's programs pull data from many lines and sites into centralized or cloud platforms, and AI agents increasingly act on what they find.

Responsible programs keep a person in the loop and require explicit approval before AI changes anything physical. That approval step protects the plant from a bad recommendation. It can't protect the connection the AI depends on, because anyone who reaches that connection can use it without asking.

Most AI data paths are built like remote access

Trace how plant data reaches an AI platform, and you'll usually find an inbound port open to a server or a jump-host in the industrial DMZ, a site-to-site VPN into a cloud environment, a vendor-managed gateway with its own route back to the OEM, or long-lived API keys tied to a shared service account. Each of these works, and each one leaves something listening that a scanner can find.

Picture a loading dock door propped open all night because a delivery truck arrives every morning. The truck gets in easily, and so does anyone else who notices the gap. More than 30 Minnesota water systems learned this when attackers reached their internet-exposed PLCs directly. An AI pipeline built on the same patterns props open another door, and it usually reaches more of the plant than any single vendor connection.

Give each AI use case only the reach it needs

Plants already know how to do this. Engineers who scope a new conduit under IEC 62443 apply the same discipline an AI connection needs, and the work doesn't have to slow the project down. Five practices make the biggest difference:

  • Start with the use case. Predictive maintenance needs sensor and historian data, and it has no reason to reach a PLC.

  • Scope each connection to one process and one destination.

  • Separate the path that carries telemetry out from any path that sends changes in. When AI recommends a change, route it through the operator's existing workflow and change management.

  • Retire standing credentials and tunnels that stay open between jobs.

  • Start connections from inside the plant, so nothing in the OT network waits for inbound traffic.

Data diodes are a strong option for pure telemetry because they enforce one-way flow in hardware. That strength is also their limit, since nothing comes back at all. Use cases that need acknowledgments, configuration pulls, or vendor support call for a different approach.

How Terniion connects plant data without exposing the plant

Terniion puts these practices in place without a network redesign. STLink runs at the plant and at the AI platform, and both sides connect outbound. The Connector joins the two halves without decrypting the traffic, so the OT network stays non-addressable from any outside vantage point. Each tunnel links one approved process to one destination, which means an ingestion process reaches the historian it's approved for and nothing else. Commander manages that policy centrally.

Legacy equipment can stay as it is. STLink runs on an adjacent supported host, so controllers that can't take an agent don't need one. 

The next AI connection doesn't have to be the next way in

AI belongs in OT, and the plants that get the most from it will connect more systems, sites, and data sources over the next few years. Each of those connections will either be something an attacker can find or something they can't. Build the data path as if it will carry commands someday. For a growing number of plants, it already does.

Frequently asked questions

Does AI in OT require an internet connection to the plant?

Not always, since edge AI can run entirely on site. Cloud analytics need data to leave the plant, and outbound-only connections let that happen without exposing anything inside the OT network to inbound traffic.

Is a data diode enough to secure OT data for AI?

For one-way telemetry, a data diode is a strong control. It can't support use cases that need return traffic, so many plants pair one-way paths for telemetry with narrowly scoped, process-level connections for everything else.

What OT data does predictive maintenance need?

Predictive maintenance usually needs sensor readings, historian data, and maintenance records. It rarely needs direct access to controllers, which makes it a strong first use case for tightly scoped connections.


Related content

Visibility and control at the industrial edge | Xiid + Portainer
White Paper

Visibility and control at the industrial edge | Xiid + Portainer

How Portainer and Xiid Terniion unify visibility across distributed OT infrastructure; self-hosted, outbound-only, with zero inbound exposure.

Read White Paper